empulse/KiCAD-MCP-Qwen3.5-4B-GGUF

KiCAD-MCP-Qwen3.5-4B

A fine-tuned Qwen3.5-4B model specialized for KiCAD PCB design assistance via the Model Context Protocol (MCP). Trained to select and invoke the correct tools from a 159-tool KiCAD MCP server covering schematic capture, PCB layout, DRC, export, library management, and more.

Capabilities

• Tool Selection: Correctly identifies which of 159 KiCAD MCP tools to call based on natural language requests
• Parameter Extraction: Extracts component references, coordinates, net names, and other parameters from user requests
• Multi-Step Reasoning: Plans sequences of tool calls for complex PCB design tasks
• Error Handling: Interprets tool error responses and suggests fixes
• Refusal: Declines off-topic or destructive requests appropriately
• Domain Knowledge: Understands PCB design concepts (DRC rules, clearances, layer stacks, component placement)

Tool Categories

The model was trained on a KiCAD MCP server with tools across these categories:

Category	Tools	Examples
Project	5	create_project, open_project, save_project
Schematic	22+	add_schematic_component, add_wire, connect_to_net, annotate_schematic
Board	12	set_board_size, add_layer, add_board_outline, get_board_2d_view
Component	17	place_component, move_component, rotate_component, find_component
DRC	8	run_drc, set_design_rules, get_drc_violations, check_clearance
Export	8	export_gerber, export_pdf, export_bom, export_3d
Routing	14	route_trace, add_via, add_copper_pour, route_differential_pair
Library	8	search_footprints, get_footprint_info, list_libraries
And more	...	Symbol creation, JLCPCB integration, freerouting, UI management

See tool_schema.json for the complete 159-tool schema with descriptions and parameter definitions.

Usage

With llama.cpp

# Download the model
huggingface-cli download empulse/KiCAD-MCP-Qwen3.5-4B-GGUF --local-dir ./model

# Run with llama-server (requires Qwen3.5 support — llama.cpp build from Feb 2026+)
llama-server \
  -m model/kicad-mcp-4b-sft-v3-bf16.gguf \
  -ngl 99 \
  --cache-type-k bf16 --cache-type-v bf16 \
  -c 4096 \
  --port 8080

# Query via OpenAI-compatible API
curl http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "messages": [
      {"role": "system", "content": "You are a KiCAD PCB design assistant with MCP tools."},
      {"role": "user", "content": "Run a DRC check on the board"}
    ],
    "temperature": 0.1
  }'

Important Notes

• GGUF format: bf16 (full precision, 8.4GB). Quantize with llama-quantize if needed.
• KV cache must be bf16: Use --cache-type-k bf16 --cache-type-v bf16 with llama.cpp. The Gated Delta Network architecture produces incorrect results with f16 cache.
• Qwen3.5 support required: llama.cpp builds from February 2026+ include Qwen3.5 GDN support (PR #19435).
• Thinking mode: The model uses <think>...</think> blocks for reasoning by default. Use /nothink prefix to disable.

Training

• Base model: Qwen/Qwen3.5-4B
• Architecture: Gated Delta Network (GDN) — hybrid attention, not standard transformer
• Method: LoRA (rank 32, alpha 64) with two-phase training
  • Phase 1: Continual Pre-Training on domain-specific corpora (MCP protocol, KiCAD documentation, PCB design knowledge)
  • Phase 2: Supervised Fine-Tuning on tool-calling examples
• Framework: ms-swift 4.0.2 with DeepSpeed ZeRO-2
• Hardware: 2x NVIDIA RTX 3090 (48GB total)
• Precision: bfloat16 throughout (required for GDN architecture)

Evaluation

Category	Score
Tool Selection (10 tests)	10/10
No-Tool Response (2 tests)	2/2
Refusal (1 test)	1/1
Domain Knowledge (1 test)	1/1
Total	14/14 (100%)

Limitations

• Trained primarily on a specific KiCAD MCP server implementation — tool names and schemas may differ from other servers
• 4B parameter model — complex multi-step reasoning chains may be less reliable than larger models
• Domain knowledge is focused on KiCAD v10 and common PCB design patterns
• Not trained for vision/image tasks despite Qwen3.5's multimodal capabilities

License

Apache 2.0